CN102629318B - Fingerprint image segmentation method based on support vector machine - Google Patents

Abstract

The invention discloses a fingerprint image segmentation method based on a support vector machine (SVM), which comprises the following steps: firstly, carrying out primary coarse segmentation on a fingerprint image by utilizing gradient features and removing most of background regions; then extracting the contrast ratio, the direction consistency, the Gabor filter variance deviation and other features of a fingerprint foreground region based on coarse segmentation and dividing the foreground region obtained by the coarse segmentation into an effective foreground region and an invalid foreground region by utilizing the support vector machine; and finally, carrying out morphological post-processing on the segmentation results. The fingerprint image segmentation method has the advantages that not only can the background regions with high noise be segmented, but also fuzzy regions with bad quality and unrecoverable textures can be separated from one another at the same time, so that the post-processing is only required to be carried out aiming at the effective background region to provide a favorable help for subsequent fingerprint quality evaluation and fingerprint pre-processing; and the segmentation accuracy is higher, and the adaptability is stronger.

Description

A kind of fingerprint image dividing method based on support vector machine

Technical field

The present invention relates to a kind of fingerprint image dividing method based on support vector machine, belong to the fingerprint identification technology field.

Background technology

Fingerprint because it has uniqueness, permanent and stable, makes fingerprint recognition more and more become the biological identification technology of the main flow of current identification application as the key character of human body.It is the important component part of fingerprint recognition system that fingerprint image is cut apart.The fingerprint dividing processing can make subsequent treatment concentrate on effective preceding scenic spot to carry out, thereby reduces calculated amount effectively, reduces the time of fingerprint image preprocessing, can also improve the degree of accuracy of feature extraction simultaneously, reduces the fingerprint storage space.

At present, how to cut apart the research emphasis that the inferior quality fingerprint image is automatic fingerprint recognition field.Such as in the actual acquisition process, because the influence of collection head surface, light and collecting device itself, contain a large amount of noises in the feasible image background that collects, these noises make that the gray-scale statistical characteristics of the background area in the fingerprint image is similar to finger-print region, be divided into prospect mistakenly easily, to cause in characteristic extraction procedure, extracting a lot of false minutiae point, reduce discrimination.And because the influence of factors such as the dynamics of pressing of the clean level of fingerprint skin, fingerprint, noise, make some real textured region poor quality, be difficult in subsequent treatment recover clearly that lines comes.If irrecoverable zone is excessive in the true lines of fingerprint, this fingerprint should be gathered again.

Summary of the invention

The objective of the invention is to overcome the deficiencies in the prior art, a kind of fingerprint image dividing method of the classification based on support vector machine (SVM) is provided, can not only be partitioned into the strong noise background area in the fingerprint image, simultaneously can remove the irrecoverable zone of lines in the true finger-print region, for follow-up fingerprint quality evaluation and fingerprint pre-service provide favourable help.

According to technical scheme provided by the invention, described fingerprint image dividing method based on support vector machine comprises following steps:

Step 1, with fingerprint image be divided into non-overlapping copies, size is 3 * 3 fritter, as the internal layer piece of cutting apart, uses W _In(i, j) fritter of the capable j row of expression i;

Step 2, utilize the gradient feature of image, fingerprint image is carried out coarse segmentation, obtain coarse segmentation image mask, concrete steps are as follows:

2.1, the edge pixel of mark ridge, paddy: (x is y) along x direction of principal axis and the axial gradient vector G of y for pixel I in the calculated fingerprint image _x(x, y) and G _y(x, y), if | G _x(x, y) | 〉=G _ThreAnd | G _y(x, y) | 〉=G _Thre, then P (x y)=1, represents that this pixel is the edge pixel of ridge, paddy, otherwise P (x, y)=0, wherein, G _ThreBe Grads threshold;

2.2, judge that according to the degree of rarefication of edge pixel each fritter in the image is foreground blocks or background piece: getting a size is b * b, geometric center and W _In(the outer piece as cutting apart is designated as W for i, the j) image block of geometric center coincidence _{Out_1}(i, j); Statistical picture piece W _{Out_1}(i, the j) number of middle edge pixel is as internal layer piece W _In(i, j) degree of rarefication of middle edge pixel

$N_{W_{\mathrm{in}}(i,j)}=\underset{(x,y)\∈W_{\mathrm{out}\_1}(i,j)}{\mathrm{\Σ}}P(x,y)---\left(1\right)$

If Then this image block is foreground blocks,

Otherwise this image block is the background piece,

N _ThreBe the degree of rarefication threshold value;

Step 3, coarse segmentation image mask is carried out aftertreatment, obtain preliminary segmentation result, concrete steps are as follows:

3.1, for each foreground blocks

If the number of foreground blocks is less than 4 in its eight neighborhood, then this foreground blocks is labeled as the background piece

3.2, for each background piece

If the number of foreground blocks is more than or equal to 4 in its eight neighborhood, then this background piece is labeled as foreground blocks

3.3, repeated execution of steps 3.2, until there not being the background piece to be marked as foreground blocks;

Step 4, on the basis of coarse segmentation image mask, if

Extract its contrast, direction consistance, Gabor filtering variance deviation, utilize support vector machine to classify, cut the result thereby obtain segmentation.

If 4.1

Getting a size is c * c, geometric center and W _In(i, j) the outer piece W of the image of geometric center coincidence _{Out_2}(i j), extracts W _{Out_2}(i, contrast j), direction consistance, Gabor filtering variance deviation are as internal layer piece W _In(i, feature j) specifically comprises:

A) calculate contrast con (i, j):

$m(i,j)=\frac{s(i,j)}{n(i,j)}---\left(2\right)$

$d(i,j)=\frac{s_1(i,j)}{n_1(i,j)}-\frac{s_2(i,j)}{n_2(i,j)}---\left(3\right)$

$\mathrm{con}(i,j)=\frac{d(i,j)}{m(i,j)}---\left(4\right)$

Wherein, n (i, j) and s (i j) is W in the image block _{Out_2}(i, j) number of all pixels and gray-scale value addition and, n ₁(i, j) and s ₁(i j) is image block W _{Out_2}(i, j) in all gray-scale values greater than m (i, the number of pixel j) and gray-scale value addition and, n ₂(i, j) and s ₂(i j) is image block W _{Out_2}(i, j) in all gray-scale values less than m (i, the number of pixel j) and gray-scale value addition and;

B) calculated direction consistance coh (i, j):

$\mathrm{coh}(i,j)=\frac{{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)-G_y^2(x,y))\right)}^2+{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(2\·G_x(x,y)\·G_y(x,y))\right)}^2}{{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)+G_y^2(x,y))\right)}^2}---\left(5\right)$

C) calculate Gabor filtering feature variance deviation:

Employing has the two-dimentional even Gabor wave filter of directional selectivity and frequency selectivity, and expression formula is as follows

$h(x,y,\mathrm{\φ},f)=\exp \{-\frac{1}{2}(\frac{x_{\mathrm{\φ}}^2}{{\mathrm{\δ}}_x^2}+\frac{y_{\mathrm{\φ}}^2}{{\mathrm{\δ}}_y^2})\}\cos \left(2\mathrm{\πf}{x}_{\mathrm{\φ}}\right)---\left(6\right)$

Wherein, x _φ=xcos φ+ysin φ, y _φ=-xsin φ+ycos φ, φ are the direction of Gabor wave filter, and f is the filter center frequency, are made as the average frequency value of fingerprint ridge line here; δ _xAnd δ _yBe respectively along the Gaussian envelope constant of x axle and y axle, these two parameters are all got 0.4; To image block W _{Out_2}(i, j) carry out the Gabor filtering of both direction:

$I_{W_{\mathrm{out}\_2}(i,j)}^{\′}(x,y)=\underset{u=-R}{\overset{R}{\mathrm{\Σ}}}\underset{v=-R}{\overset{R}{\mathrm{\Σ}}}h(u,v,\mathrm{\θ}(i,j),f)\·I_{W_{\mathrm{out}\_2}(i,j)}(x-u,y-v)---\left(7\right)$

$I_{W_{\mathrm{out}\_2}(i,j)}^{\′\′}(x,y)=\underset{u=-R}{\overset{R}{\mathrm{\Σ}}}\underset{v=-R}{\overset{R}{\mathrm{\Σ}}}h(u,v,\mathrm{\θ}(i,j)+\frac{\mathrm{\π}}{2},f)\·I_{W_{\mathrm{out}\_2}(i,j)}(x-u,y-v)---\left(8\right)$

Wherein, R represents the radius of wave filter, and (i j) is image block W to θ _{Out_2}(i, texture principal direction j),

Be image block W _{Out_2}(i, time direction of texture j), θ (i, computing formula j) is as follows:

$\mathrm{\θ}(i,j)=\frac{1}{2}{\tan }^{-1}\frac{\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)-G_y^2(x,y))}{\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x(x,y)\·G_y(x,y))}---\left(9\right)$

Gabor filtering variance deviation Gabor_dif (i, j):

$\mathrm{Gabor}\_1=\frac{1}{\mathrm{num}}\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}{(I_{W_{\mathrm{out}\_2}(i,j)}^{\′}(x,y)-m_1(i,j))}^2---\left(10\right)$

$\mathrm{Gabor}\_2=\frac{1}{\mathrm{num}}\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}{(I_{W_{\mathrm{out}\_2}(i,j)}^{\′\′}(x,y)-m_2(i,j))}^2---\left(11\right)$

Gabor_dif(i，j)＝|Gabor_1-Gabor_2| (12)

Wherein, num is the number of pixels in the image block, m ₁(i, j) and m ₂(i j) is respectively the pixel average of image block after the filtering of texture primary and secondary both direction;

Obtain the proper vector of image subblock thus, consider the difference between the different fingerprint images in the same fingerprint base, proper vector is carried out normalized;

4.2, the structure training sample, training SVM: the fingerprint image that will be used for training at first carries out coarse segmentation, the proper vector of foreground area among the computed segmentation result, and hand picking goes out partial noise zone and the irrecoverable zone of lines then, as the inactive area sample, be designated as [con (i, j), coh (i, j), Gabor_dif (i, j) ,-1]; Pick out clear of part fingerprint lines and marked, as effective foreground area sample, be designated as [con (i, j), coh (i, j), Gabor_dif (i, j) ,+1]; Utilize training sample that SVM is trained, obtain final SVM model;

4.3, revise the mask image with the SVM that trains: to each image block, if

Then calculate its contrast, direction consistance and Gabor filtering variance deviation, obtain proper vector, the SVM that input trains; The output of the correspondence of SVM if+1, then declaring this piece is effective foreground area,

If-1, then declare the invalid foreground area of this piece,

Result after step 5, employing morphological method are cut segmentation carries out aftertreatment.

Further, described step 2 ectomesoderm block size b=32, Grads threshold G _Thre=8, degree of rarefication threshold value N _Thre=200.

Further, extract the outer block size c=16 of feature in the described step 4, the kernel function of SVM is selected the radially basic kernel function of Gauss.

Described step 5 adopts morphological method to carry out aftertreatment, owing to be subjected to noise effect, discrete foreground blocks and background piece can occur in the segmentation result; For each foreground blocks, if the number of foreground blocks is less than 4 in its eight neighborhood, then this foreground blocks is labeled as the background piece; For each background piece, if the number of background piece is less than 4 in its eight neighborhood, then this piece is labeled as foreground blocks.

Beneficial effect of the present invention: 1) can remove a large amount of low noises background area rapidly and accurately.Though it is bigger that the background area, also has some some Grad owing to The noise, yet the distribution of these pixels is generally all comparatively sparse, so calculating pixel point degree of rarefication within the specific limits just can separate foreground area and background area preferably.2) utilize svm classifier can remove the strong noise zone and lines can not recover inactive area such as zone, obtain effective finger-print region of clean mark.The removal of these inactive area is for follow-up fingerprint recognition system provides favourable help.For example: if irrecoverable zone is excessive in the true lines of fingerprint, disallowable in then can follow-up fingerprint quality evaluation procedure, this fingerprint be gathered again.

Description of drawings

Fig. 1 be image block internal layer piece piece outer with it concern synoptic diagram.

Fig. 2 is overview flow chart of the present invention.

Embodiment

Below in conjunction with accompanying drawing embodiment of the present invention are elaborated.

Fig. 1 has described the geometric relationship of image internal layer piece and outer piece.As the internal layer piece W of cutting apart _In(i, j), its size is a * a, W _Out(i is its corresponding outer layer piece j), is that size is b * b, geometric center and W _In(i, the j) image block of geometric center coincidence, and b is greater than a.

As shown in Figure 2, idiographic flow of the present invention is as follows.

1) fingerprint image is divided into non-overlapping copies, size is 3 * 3 fritter, as the internal layer piece of cutting apart, uses W _In(i, j) fritter of the capable j row of expression i.In actual applications, the size of internal layer piece can be got different values according to the difference of the resolution of fingerprint image.For being without loss of generality, the size of internal layer piece is made as 3 * 3 here.

2) utilize the gradient feature of image, fingerprint image is carried out coarse segmentation, obtain coarse segmentation image mask.Coarse segmentation is with removing the less background of noise in the background, to save the calculated amount of subsequent step.Concrete steps are as follows:

The edge pixel of mark ridge, paddy.(x is y) along x direction of principal axis and the axial gradient vector G of y for pixel I in the calculated fingerprint image I _x(x, y) and G _y(x, y).If | G _x(x, y) | 〉=G _ThreAnd | G _y(x, y) | 〉=G _Thre, then (x y)=1, represents that this pixel is the edge pixel of ridge, paddy to P; Otherwise P (x, y)=0.Wherein, G _ThreBe Grads threshold.

Judge that according to the degree of rarefication of edge pixel each fritter in the image is foreground blocks or background piece.Getting a size is b * b, geometric center and W _In(the outer piece as cutting apart is designated as W for i, the j) image block of geometric center coincidence _{Out_1}(i, j).Statistical picture piece W _{Out_1}(i, the j) number of middle edge pixel is as internal layer piece W _In(i, j) degree of rarefication of middle edge pixel:

$N_{W_{\mathrm{in}}(i,j)}=\underset{(x,y)\∈W_{\mathrm{out}\_1}(i,j)}{\mathrm{\Σ}}P(x,y)---\left(1\right)$

If

Then this image block is foreground blocks,

Otherwise this image block is the background piece,

In order to make degree of rarefication have stronger noise immunity, the outer block size here is 32 * 32, degree of rarefication threshold value N _Thre=200.Grads threshold arranges less here, G _Thre=8.

3) coarse segmentation image mask is carried out aftertreatment, obtain preliminary segmentation result.

3.1 for each foreground blocks

If the number of foreground blocks is less than 4 in its eight neighborhood, then this foreground blocks is labeled as the background piece

3.2, for each background piece

If the number of foreground blocks is more than or equal to 4 in its eight neighborhood, then this background piece is labeled as foreground blocks

3.3, repeated execution of steps 3.2, until there not being the background piece to be marked as foreground blocks;

Through primary coarse segmentation, some is done partially or wets to such an extent that the gray scale between finger-print region ridge, the paddy is comparatively approaching partially, and the edge is not obvious, is divided into background easily, and in fact these zones may be that lines can recover the zone, and the foreground area of fingerprint is included in these zones again after the aftertreatment.But some dead space image block is also by the effective district of being divided into of mistake.Some fractures in the serious zone of noise pollution, the real lines are serious such as being subjected in the background, the zone that is difficult to recover lines etc.These zones will be cut apart away them in secondary splitting.

4) on the basis of preliminary split image mask, if

Extract its features such as contrast, direction consistance, Gabor filtering variance deviation, utilize support vector machine to classify, cut the result thereby obtain segmentation.

If 4.1

Get a size and be 16 * 16, geometric center and W _In(i, j) the outer piece W of the image of geometric center coincidence _{Out_2}(i, j).Extract W _{Out_2}(i, features such as contrast j), direction consistance, Gabor filtering variance deviation are as internal layer piece W _In(i, feature j).

A) contrast con (i, j):

$m(i,j)=\frac{s(i,j)}{n(i,j)}---\left(2\right)$

$d(i,j)=\frac{s_1(i,j)}{n_1(i,j)}-\frac{s_2(i,j)}{n_2(i,j)}---\left(3\right)$

$\mathrm{con}(i,j)=\frac{d(i,j)}{m(i,j)}---\left(4\right)$

Wherein, n (i, j) and s (i j) is W in the image block _{Out_2}(i, j) number of all pixels and gray-scale value addition and.n ₁(i, j) and s ₁(i j) is image block W _{Out_2}(i, j) in all gray-scale values greater than m (i, the number of pixel j) and gray-scale value addition and.n ₂(i, j) and s ₂(i j) is image block W _{Out_2}(i, j) in all gray-scale values less than m (i, the number of pixel j) and gray-scale value addition and.

B) direction consistance coh (i, j):

$\mathrm{coh}(i,j)=\frac{{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)-G_y^2(x,y))\right)}^2+{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(2\·G_x(x,y)\·G_y(x,y))\right)}^2}{{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)+G_y^2(x,y))\right)}^2}---\left(5\right)$

C) Gabor filtering feature variance deviation:

Employing has the two-dimentional even Gabor wave filter of directional selectivity and frequency selectivity, and its expression formula is as follows:

$h(x,y,\mathrm{\φ},f)=\exp \{-\frac{1}{2}(\frac{x_{\mathrm{\φ}}^2}{{\mathrm{\δ}}_x^2}+\frac{y_{\mathrm{\φ}}^2}{{\mathrm{\δ}}_y^2})\}\cos \left(2\mathrm{\π}{x}_{\mathrm{\φ}}\right)---\left(6\right)$

x _φ＝xcosφ+ysinφ (7)

y _φ＝-xsinφ+ycosφ (8)

Wherein, φ is the direction of Gabor wave filter, and f is the filter center frequency, δ _xAnd δ _yBe respectively along the Gaussian envelope constant of x axle and y axle.Image block is carried out the Gabor filtering of texture principal direction and time direction:

$I_{W_{\mathrm{out}\_2}(i,j)}^{\′}(x,y)=\underset{u=-R}{\overset{R}{\mathrm{\Σ}}}\underset{v=-R}{\overset{R}{\mathrm{\Σ}}}h(u,v,\mathrm{\θ}(i,j),f)\·I_{W_{\mathrm{out}\_2}(i,j)}(x-u,y-v)---\left(9\right)$

$I_{W_{\mathrm{out}\_2}(i,j)}^{\′\′}(x,y)=\underset{u=-R}{\overset{R}{\mathrm{\Σ}}}\underset{v=-R}{\overset{R}{\mathrm{\Σ}}}h(u,v,\mathrm{\θ}(i,j)+\frac{\mathrm{\π}}{2},f)\·I_{W_{\mathrm{out}\_2}(i,j)}(x-u,y-v)---\left(10\right)$

Wherein, centre frequency f is made as the average frequency of fingerprint ridge line, and the present invention gets 0.1.δ _xAnd δ _yValue 0.4.(i j) is image block W to θ _{Out_2}(i, texture principal direction j),

Be image block W _{Out_2}(i, inferior direction j).The computing formula of texture principal direction is as follows:

$\mathrm{\θ}(i,j)=\frac{1}{2}{\tan }^{-1}\frac{\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)-G_y^2(x,y))}{\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x(x,y)\·G_y(x,y))}---\left(11\right)$

Gabor filtering variance deviation Gabor_dif (i, j):

$\mathrm{Gabor}\_1=\frac{1}{\mathrm{num}}\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}{(I_{W_{\mathrm{out}\_2}(i,j)}^{\′}(x,y)-m_1(i,j))}^2---\left(12\right)$

$\mathrm{Gabor}\_2=\frac{1}{\mathrm{num}}\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}{(I_{W_{\mathrm{out}\_2}(i,j)}^{\′\′}(x,y)-m_2(i,j))}^2---\left(13\right)$

Gabor_dif(i，j)＝|Gabor_1-Gabor_2| (14)

Wherein, num is image block W _{Out_2}(i, j) Nei number of pixels, m ₁(i, j) and m ₂(i j) is respectively the pixel average of image block after principal direction and time direction filtering.

Can obtain the proper vector of image block thus, consider the difference between the different fingerprint images in the same fingerprint base, proper vector is carried out normalized.

4.2 the structure training sample, training SVM.The fingerprint image that will be used for training at first carries out coarse segmentation, the proper vector of foreground area among the computed segmentation result, and hand picking goes out partial noise zone and the irrecoverable zone of lines then, as the inactive area sample, be designated as [con (i, j), coh (i, j), Gabor_dif (i, j) ,-1].Pick out clear of part fingerprint lines and marked, as effective foreground area sample, be designated as [con (i, j), coh (i, j), Gabor_dif (i, j) ,+1].The kernel function of SVM is selected the radially basic kernel function of Gauss, utilizes training sample that SVM is trained, and obtains final SVM model.

4.3 revise the mask image with the SVM that trains.To each image block, if

Calculate its contrast, direction consistance and Gabor filtering variance deviation, obtain proper vector, the SVM that input trains.The output of the correspondence of SVM if+1, then declaring this piece is effective foreground area,

If-1, then declaring this piece is invalid foreground area,

5) adopt the result after morphological method is cut segmentation to carry out aftertreatment.Some isolated foreground blocks and background pieces can appear in the segmentation result.For each foreground blocks, if the number of foreground blocks is less than 4 in its eight neighborhood, then this foreground blocks is labeled as the background piece.For each background piece, if the number of background piece is less than 4 in its eight neighborhood, then this piece is labeled as foreground blocks.

Fingerprint image is cut apart gray-scale statistical information and the texture information that choosing of feature utilized image block among the present invention.Contrast is relatively more responsive to grey scale change.The direction consistance has reflected the direction degree of consistency of having a few in the piece.The Gabor wave filter has the empty resolution frequently of optimum associating, is equivalent to image block has been carried out the filtering enhancing at this.The effective preceding scenic spot of fingerprint image streakline is light and dark, and grey scale change is violent, has the grain direction of rule, and through after the Gabor filtering of both direction, it is obvious that principal direction strengthens effect, so the deviation of filtering variance is bigger; And the intensity profile in strong noise background area and the irrecoverable zone of lines is even relatively, and the direction degree of consistency is low, and principal direction is all not obvious with the enhancing effect of time direction, so the deviation of filtering variance is less.

Claims (4)

1. the fingerprint image dividing method based on support vector machine is characterized in that, comprises following steps:

Step 1, with fingerprint image be divided into non-overlapping copies, size is 3 * 3 fritter, as the internal layer piece of cutting apart, uses W _In(i, j) fritter of the capable j row of expression i;

Step 2, utilize the gradient feature of image, fingerprint image is carried out coarse segmentation, obtain coarse segmentation image mask, concrete steps are as follows:

2.1, the edge pixel of mark ridge, paddy: (x is y) along x direction of principal axis and the axial gradient vector G of y for pixel I in the calculated fingerprint image _x(x, y) and G _y(x, y), if | G _x(x, y) | 〉=G _ThreAnd | G _y(x, y) | 〉=G _Thre, then P (x y)=1, represents that this pixel is the edge pixel of ridge, paddy, otherwise P (x, y)=0, wherein, G _ThreBe Grads threshold;

2.2, judge that according to the degree of rarefication of edge pixel each fritter in the image is foreground blocks or background piece: getting a size is b * b, geometric center and W _In(the outer piece as cutting apart is designated as W for i, the j) image block of geometric center coincidence _{Out_1}(i, j); Statistical picture piece W _{Out_1}(i, the j) number of middle edge pixel is as internal layer piece W _In(i, j) degree of rarefication of middle edge pixel

$N_{W_{\mathrm{in}}(i,j)}=\underset{(x,y)\∈W_{\mathrm{out}\_1}(i,j)}{\mathrm{\Σ}}P(x,y)---\left(1\right)$

If

Then this image block is foreground blocks,

Otherwise this image block is the background piece,

N _ThreBe the degree of rarefication threshold value;

Step 3, coarse segmentation image mask is carried out aftertreatment, obtain preliminary segmentation result, concrete steps are as follows:

3.1, for each foreground blocks

If the number of foreground blocks is less than 4 in its eight neighborhood, then this foreground blocks is labeled as the background piece

3.2, for each background piece

If the number of foreground blocks is more than or equal to 4 in its eight neighborhood, then this background piece is labeled as foreground blocks

3.3, repeated execution of steps 3.2, until there not being the background piece to be marked as foreground blocks;

Step 4, on the basis of coarse segmentation image mask, if

Extract its contrast, direction consistance, Gabor filtering variance deviation, utilize support vector machine to classify, cut the result thereby obtain segmentation:

If 4.1

Getting a size is c * c, geometric center and W _In(i, j) the outer piece W of the image of geometric center coincidence _{Out_2}(i j), extracts W _{Out_2}(i, contrast j), direction consistance, Gabor filtering variance deviation are as internal layer piece W _In(i, feature j) specifically comprises:

A) calculate contrast con (i, j):

$m(i,j)=\frac{s(i,j)}{n(i,j)}---\left(2\right)$

$d(i,j)=\frac{s_1(i,j)}{n_1(i,j)}-\frac{s_2(i,j)}{n_2(i,j)}---\left(3\right)$

$\mathrm{con}(i,j)=\frac{d(i,j)}{m(i,j)}---\left(4\right)$

Wherein, n (i, j) and s (i j) is W in the image block _{Out_2}(i, j) number of all pixels and gray-scale value addition and, n ₁(i, j) and s ₁(i j) is image block W _{Out_2}(i, j) in all gray-scale values greater than m (i, the number of pixel j) and gray-scale value addition and, n ₂(i, j) and s ₂(i j) is image block W _{Out_2}(i, j) in all gray-scale values less than m (i, the number of pixel j) and gray-scale value addition and;

B) calculated direction consistance coh (i, j):

$\mathrm{coh}(i,j)=\frac{{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)-G_y^2(x,y))\right)}^2+{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(2\·G_x(x,y)\·G_y(x,y))\right)}^2}{{\left(\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)+G_y^2(x,y))\right)}^2}---\left(5\right)$

C) calculate Gabor filtering feature variance deviation:

Employing has the two-dimentional even Gabor wave filter of directional selectivity and frequency selectivity, and expression formula is as follows

$h(x,y,\mathrm{\φ},f)=\exp \{-\frac{1}{2}(\frac{x_{\mathrm{\φ}}^2}{{\mathrm{\δ}}_x^2}+\frac{y_{\mathrm{\φ}}^2}{{\mathrm{\δ}}_y^2})\}\cos \left(2\mathrm{\π}{\mathrm{fx}}_{\mathrm{\φ}}\right)---\left(6\right)$

Wherein, x _φ=xcos φ+ysin φ, y _φ=-xsin φ+ycos φ, φ are the direction of Gabor wave filter, and f is the filter center frequency, are made as the average frequency value of fingerprint ridge line here; δ _xAnd δ _yBe respectively along the Gaussian envelope constant of x axle and y axle, these two parameters are all got 0.4; Image block is carried out the Gabor filtering of both direction:

$I_{W_{\mathrm{out}\_2}(i,j)}^{\′}(x,y)=\underset{u=-R}{\overset{R}{\mathrm{\Σ}}}\underset{v=-R}{\overset{R}{\mathrm{\Σ}}}h(u,v,\mathrm{\θ}(i,j),f)\·I_{W_{\mathrm{out}\_2}(i,j)}(x-u,y-v)---\left(7\right)$

$I_{W_{\mathrm{out}\_2}(i,j)}^{\′\′}(x,y)=\underset{u=-R}{\overset{R}{\mathrm{\Σ}}}\underset{v=-R}{\overset{R}{\mathrm{\Σ}}}h(u,v,\mathrm{\θ}(i,j)+\frac{\mathrm{\π}}{2},f)\·I_{W_{\mathrm{out}\_2}(i,j)}(x-u,y-v)---\left(8\right)$

Wherein, R represents the radius of wave filter, and (i j) is image block W to θ _{Out_2}(i, texture principal direction j),

Be image block W _{Out_2}(i, time direction of texture j), θ (i, computing formula j) is as follows:

$\mathrm{\θ}(i,j)=\frac{1}{2}{\tan }^{-1}\frac{\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x^2(x,y)-G_y^2(x,y))}{\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}(G_x(x,y)\·G_y(x,y))}---\left(9\right)$

Gabor filtering variance deviation Gabor_dif (i, j):

$\mathrm{Gabor}\_1=\frac{1}{\mathrm{num}}\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}{(I_{W_{\mathrm{out}\_2}(i,j)}^{\′}(x,y)-m_1(i,j))}^2---\left(10\right)$

$\mathrm{Gabor}\_2=\frac{1}{\mathrm{num}}\underset{(x,y)\∈W_{\mathrm{out}\_2}(i,j)}{\mathrm{\Σ}}{(I_{W_{\mathrm{out}\_2}(i,j)}^{\′\′}(x,y)-m_2(i,j))}^2---\left(11\right)$

Gabor_dif(i,j)=|Gabor_1-Gabor_2| （12）

Wherein, num is image block W _{Out_2}(i, j) Nei number of pixels, m ₁(i, j) and m ₂(i j) is respectively the pixel average of image block after principal direction and time direction filtering;

Obtain the proper vector of image subblock thus, consider the difference between the different fingerprint images in the same fingerprint base, proper vector is carried out normalized;

4.2, the structure training sample, training SVM: the fingerprint image that will be used for training at first carries out coarse segmentation, the proper vector of foreground area among the computed segmentation result, and hand picking goes out partial noise zone and the irrecoverable zone of lines then, as the inactive area sample, be designated as [con (i, j), coh (i, j), Gabor_dif (i, j) ,-1]; Pick out clear of part fingerprint lines and marked, as effective foreground area sample, be designated as [con (i, j), coh (i, j), Gabor_dif (i, j) ,+1]; Utilize training sample that SVM is trained, obtain final SVM model;

4.3, revise the mask image with the SVM that trains: to each image block, if

Then calculate its contrast, direction consistance and Gabor filtering variance deviation, obtain proper vector, the SVM that input trains; The output of the correspondence of SVM if+1, then declaring this piece is effective foreground area,

If-1, then declare the invalid foreground area of this piece,

Result after step 5, employing morphological method are cut segmentation carries out aftertreatment.

2. a kind of fingerprint image dividing method based on support vector machine according to claim 1 is characterized in that, described step 2 ectomesoderm block size b=32, Grads threshold G _Thre=8, degree of rarefication threshold value N _Thre=200.

3. a kind of fingerprint image dividing method based on support vector machine according to claim 1 is characterized in that, extracts the outer block size c=16 of feature in the described step 4, and the kernel function of SVM is selected the radially basic kernel function of Gauss.

4. a kind of fingerprint image dividing method based on support vector machine according to claim 1 is characterized in that, described step 5 adopts morphological method to carry out aftertreatment, owing to be subjected to noise effect, discrete foreground blocks and background piece can occur in the segmentation result; For each foreground blocks, if the number of foreground blocks is less than 4 in its eight neighborhood, then this foreground blocks is labeled as the background piece; For each background piece, if the number of background piece is less than 4 in its eight neighborhood, then this piece is labeled as foreground blocks.

Images